BARD: Bayesian-Assisted Resource Discovery

1
BARD Bayesian-AssistedResource Discovery

• Fred Stann (USC/ISI)
• Joint Work With
• John Heidemann (USC/ISI)
• April 9, 2004

2
Motivation

• Problem Efficiency of Data Dissemination in
  Sensor Networks
• Data producers and data consumers must connect
  with each other
• Exhaustive search (a.k.a. flooding) required
• In lieu of meta-data or a priori knowledge
• Solution BARD uses Bayesian techniques
• Use prior distribution to limit flooding

3
Data Dissemination in Sensor Nets

• Resource Discovery
• Finding data matching some description
• Attribute Matching
• Routing
• Route Establishment
• Packet Forwarding
• Route Maintenance

4
Name-Based vs. Attribute-Based Routing

• IP Ad Hoc Routing
• Name-based routing with Resource Discovery
  layered on top (e.g. DNS, Google)
• Diffusion
• Attribute-based routing combined with
• Resource Discovery

5
Related Work

• Route Caching (DSR, AODV)
• Cached paths are refreshed as needed
• Data Centric Storage (DCS/GHT)
• Hash to location aware nodes
• Geographic Assist (GEAR)
• Greedy forwarding toward target
• Target Tracking (Spatio-Temporal Mcast)
• Predict target path and delivery zone
• Probabilistic (Gossip)
• Forwarding with fixed probability

6
Related Work Summary

• Each technique works well for a subset of the
  problem space comprised of all diffusion
  applications
• We desired a more general approach

7
Two-Phase Pull Diffusion
Sink
(could be multiple sinks)
target
Source
Additional source

• Original diffusion algorithm Intanagonwiwiat et
  al, 2000
• 1. flood interests from sink to source
• 2. flood exploratory data from source back to
  sink
• 3. reinforce preferred gradient(s) from sink to
  source (tree)
• 4. send data along reinforced gradients

8
Push Diffusion
Sink
(could be multiple sinks)
target
Source
Additional source

• Make sources active to avoid one flood NEW
• flood interests from sink to source
• 1. flood exploratory data from source back to
  sink
• 2. reinforce preferred gradient(s) from sink to
  source (tree)
• 3. send data along reinforced gradients

9
Statistical Approach

• Correlation in sensor networks
• Real-world events create patterns over time
• Implicit geography

10
Modeling Resource Discovery

• The Joint Probability Distribution (joint)
• Grows Exponentially

11
Bayesian Approach

• Combine prior probability with a sample.
• Keep track of reinforcements per attribute per
  neighbor as Conditional Probability Tables (CPTs)
• Simpler to maintain than a joint probability
  distribution.
• Current Sample
• Set of attributes in exploratory packet.
• Forward to high probability neighbors

12
Bayesian Approach cont
?

• Bayes requires conditional independence
• PA?N3 PA?N3?S

13
Implemented as a Diffusion Filter
The Filter Architecture in Diffusion, allows BARD
to be a selectable service.
14
BARD Filter Pre-Processing
15
BARD Filter Limited Routing
16
BARD Flooding

• Flooding When CPTs Empty
• Build up CPTs
• Periodic Flooding
• Updating CPTs in response to changing conditions
• Sliding time window
• Compensation for Hysteresis
• Low fidelity real-time events

17
BARD Simulation Experiments

• Increasing node count (and area)
• Increasing density
• Varying the number of sources
• Varying the number of sinks
• Sensitivity to transmission error
• Increasing send frequency
• Moving target

18
ns-2 Results Summary

• BARD - 28 to 78 reduction in control traffic
• BARD results improve with
• Higher node counts
• Greater node density
• Lower send rates
• BARD results are limited by
• Increased number of sources
• Dispersion of sources
• Higher send rates
• High error rates

19
Increasing Node Count Area

• Simple push overhead grows faster than BARD
• 45 ? 53 improvement in control byte overhead

20
Increasing Node Density

• Hop count doesnt increase, so efficiency
  increases
• 62 ? 73 improvement in control byte overhead

21
Complex Example

• Relative position of sources and sinks matters
• 28 ? 47 improvement in control byte overhead

22
Increasing Send Rate

• Control amortizes (convergent) with event count
• Total transmissions affected by alternate paths

23
Stayton Test Bed Experiment

• Results as expected
• Limited routing to thin side 100 by BARD
• Multiple paths on fat side
• Ns-2 simulation had qualitatively similar results

24
Ongoing Work

• More Comprehensive testbed Experiments
• Testing with limited attribute intersection
• Complete matching rules

25
Conclusions

• Applications with complex on-demand queries, and
  low data rates can benefit
• Efficiency gain is proportional to correlation of
  events over time
• Ratio of flooding to limited flooding presents a
  tradeoff of real-time response vs. efficiency
  gain